Environmental Sound Classification Using Edge Impulse and Smartphone for Tiny ML Applications

Abstract

Environmental Sound Classification (ESC) revolves around the method of identifying an audio stream that is connected to various ambient sounds. The underexplored nature of sound classification stems from overlapping sound events and noise from other sources and requires subjective human perception due to its computational complexity. This paper deals with audio classification using Edge Impulse software for real-time systems. We begin by evaluating various combinations of feature extraction methods and neural networks. Feature extraction methods, namely Mel-frequency Cepstral Coefficients (MFCC) and Mel-Frequency Energy (MFE), were applied to the Environment Sound Classification (ESC-50) audio dataset, and the resulting features were fed into Convolutional Neural Networks (CNNs) and pre-trained audio models such as YAMNet and VGGish. The study reveals that combining MFE for feature extraction with VGGish for classification yields 99% accuracy. Live classification of realtime audio data on a smartphone shows that this is an efficient method for audio classification. The real-time demonstration on a smartphone shows that classification is practical and can be performed on any edge device, anywhere.